1 / 24

A Switch-Tagged Routing Methodology for PC Clusters with VLAN Ethernet

A Switch-Tagged Routing Methodology for PC Clusters with VLAN Ethernet. 2011 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS. Authors: Michihiro Koibuchi , Member, IEEE, Tomohiro Otsuka , Tomohiro Kudoh , Member, IEEE Computer Society, and

Download Presentation

A Switch-Tagged Routing Methodology for PC Clusters with VLAN Ethernet

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Switch-Tagged RoutingMethodologyfor PC Clusters with VLAN Ethernet 2011 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS Authors: MichihiroKoibuchi, Member, IEEE, Tomohiro Otsuka, Tomohiro Kudoh, Member, IEEE Computer Society, and Hideharu Amano, Member, IEEE Computer Society Speaker: Ming-Chao Hsu, National Cheng Kung University

  2. Outline • INTRODUCTION • RELATED WORK • SWITCH-TAGGED ROUTING METHODOLOGY • FUNDAMENTAL EVALUATION • EVALUATION USING PC CLUSTERS • CONCLUSIONS

  3. INTRODUCTION • High-throughput commercial Ethernet switches are now available,and the link bandwidth of Ethernet has rapidly increased • The standardizations of 10-gigabit Ethernet(10 GbE). • As of November 2008, GbEs were employed asinterconnects on 56 percent of the TOP500 supercomputers • Recent PC clusterswith Ethernet employ system software that supports lowlatencyzero- or one-copy communication used in systemarea networks (SANs) • Most current PCclusters using Ethernet have employed simple tree-basedtopologies. • To avoid broadcast stormswhich circulate packets forever in layer-2 Ethernet • IEEE 802.1D STP or 802.1D-2004 Rapid STP (RSTP)

  4. INTRODUCTION • IEEE 802.1Q tagged virtual LAN (VLAN) technology • For setting up multiple paths between a pair of switches on loop and mesh topologies. • The existing VLAN-based routing method cannot be easily applied to most of current PC clusters • Message Passing Interface (MPI) communication libraries used in PC clusters • Usually do not support tagged VLAN technology. • Not much work exists on evaluating deadlock free routing algorithms in Ethernet, or their impact on real PC clusters with many Ethernet switches. • Switch-tagged routing methodology for PC clusters • It simply configures a switch • It disables the spanning-tree protocol (STP), allocates the VLAN sets, and optionally registers static MAC addresses of hosts for the routing. • MPI communication libraries do not need to use tagged VLAN technology • Various deadlock-free routing algorithms

  5. RELATED WORK • High-performance deterministic routing algorithms (Static MAC routing) • Statically registering MAC addresses of hosts without VLAN technology • It is difficult to stabilize the management of frames with such a configured Ethernet when a broadcast storm occurs. • Routing implementation techniques using VLAN technology (VLAN Routing) • Partitioning hosts into multiple groups in the Internet backbone for the QoS control • The VLAN technology was not intended for increasing network throughput • Multiple paths between hosts can be obtained by using VLANs( Multiple VLANs) • Each host having a different tree of the physical network • All pairs of hosts can communicate with each other via any VLAN tree topology, and there are multiple paths that consist of different link sets between each pair of hosts. • Ethernet switches support IEEE 802.1D STP or 802.1D-2004 Rapid STP (RSTP) to prevent loops in a network • STP and RSTP are not aware of VLANs. • When these protocols are enabled, all links out of a spanning tree are automatically disabled.

  6. SWITCH-TAGGED ROUTING METHODOLOGY • Frame Tagging at Switch - A switch behavior of the VLAN tagging operation • When an untagged frame enters a port, it is tagged with a default VLAN ID tag number (port VLAN ID, PVID). • Frames leaving the switch are either tagged or untagged depending on the port’s VLAN configuration. • If the port is a “tagged” member of a VLAN, the output frame is tagged with the respective VLAN ID. • If the port is an “untagged” member of a VLAN, the output frame is left untagged. Fixed VLAN assignment for the dimension-order routing (DOR) in 4 x 4 2D mesh

  7. SWITCH-TAGGED ROUTING METHODOLOGY • Renamed VLAN assignment of fat tree. • (a) Paths in fat tree

  8. SWITCH-TAGGED ROUTING METHODOLOGY • Renamed VLAN assignment of fat tree. • (b) Renamed VLAN assignment of switch s1

  9. SWITCH-TAGGED ROUTING METHODOLOGY • MAC Address Management at switches • Ethernet switches usually learn unknown MAC addresses when they receive frames. • When a path from host A to B and one from B to A use different VLANs, the intermediate switches of both paths cannot learn the destination MAC address. • This is because the MAC address self-learning procedure is independently performed on each VLAN. • This problem can be resolved through static MAC address registration. • Ethernet switches statically register pairs of MAC addresses, VLAN IDs, and output port numbers.

  10. SWITCH-TAGGED ROUTING METHODOLOGY • On/Off and Multispeed Link Regulation for Saving Power • The power consumption of links can be reduced by using the port-shutdown operation available in most commercial Ethernet switches. • Their operation was not originally intended to reduce power consumption; it is normally used to block the injection of unexpected frames from neighboring switches. • Standard management information base (MIB) + simple network management protocol (SNMP). • Power Consumption of GbE Switches (W) • The “All except ports” is the power consumption of switches when all the ports are shutdown, • “Max (port ratio)” is the power consumption when all ports are activated with 1 Gbps. • “PC5324,”“PC6224,” and “PC6248” stand for Dell PowerConnect5324(layer 2), 6224, and 6248 (layer 3) switches • “SF-420” is Planex Communications SF-0420G

  11. FUNDAMENTAL EVALUATION • Fixed VLAN assignment in fat tree (2,4,2) • switch has u upper links and d lower links, and the number of its layers is r. (u, d, r) Renamed VLAN assignment

  12. FUNDAMENTAL EVALUATION • Overhead of VLAN Operations • Dell PowerConnect 5324, 6224(layer-3), Netgear GSM7212, and Planex SF-0420G • ICMP message • The tagged VLAN operation does not affect the latency at these switches. • Latency of Switch (in Microseconds)

  13. FUNDAMENTAL EVALUATION • Overhead of VLAN Operations • Tperf 1.5 • The bandwidth of the tagged frame transfer (T-T and T-U) is thus slightly lower than that of untagged frame transfer in all switches • Renamed VLAN assignment, (U-(T)-U), has no bandwidth overhead, since it transfers untagged frames between switches. • Bandwidth of Switch (in Megabits Per Second)

  14. FUNDAMENTAL EVALUATION • Topology considered in evaluations • The “fixed path” represents frame transfers using VLAN A , • The “dynamic path” represents frame transfers using VLANs A and B that are dynamically changed at every a few seconds. • The frame bandwidth is measured at every 100 mseconds (GtrcNET-1.) • The performance of latency and throughput was unaffected even though path set is switched at every a few seconds. Performance of Path Modification (Testbed)

  15. EVALUATION USING PC CLUSTERS • Three types of PC clusters • SCorecluster - Testbed • SCore is an open-source cluster system software package that provides various parallel programming environments • Misccluster- 66-host PC cluster using six GbE switches (Dell PowerConnect 6248, 48 ports) • MiscEach switch in the Misc cluster connects to 11 hosts, • SuperNova cluster- 225-host PC cluster using the eight same GbE switches. • Each switch in the SuperNova cluster connects to 28 or 29 hosts Misc Cluster Testbed SuperNova Cluster The IEEE 802.3x link-level flow control was enabled at every port

  16. EVALUATION USING PC CLUSTERS • Evaluated indirect topologies • (a) Fat tree (2,4,2) • (b) Myri-clos(4x4). Evaluated Topologies in Testbed “Avg.H” represents the average number of switches that compose a path.

  17. EVALUATION USING PC CLUSTERS • The number of hosts was 16 in all topologies on the testbed. • maximum UDP datagram size, 1,470 bytes • UDP transfer of Tperf-1.5 • flat 1-switch network (flat, nonblocking full crossbar) in which all 16 hosts were connected to a single switch • it is hardly possible to employ such a topology in a large-scale cluster with thousands of hosts. • M-tree • Single tree-based topology with no VLANs • Network throughput • bit-reversal traffic • a host with the identifier (a0,a1,…an-1) sends a packet to the host whose identifier is the bit reversal (an-1,…,a1,a0) of the source host. • matrix transpose traffic • a host (x; y) sends a packet to the host (k-y-1; k-x-1) (k is the number of hosts in each dimension) or (k-x-1; k-y-1) when x + y =k-1.

  18. EVALUATION USING PC CLUSTERS • NAS Parallel Benchmarks(NPBs) • The unit of performance is the execution time (second) • The number of processes for parallel execution was fixed to 16 in the case of the PC-cluster testbed • the SuperNova cluster used 128 processes in CG, FT, IS, LU, and MG • And 225 processes (the maximum size) in SP and BT. • the performance of M-tree is especially degraded on FT and IS. It is known that FT and IS frequently perform the MPI_Alltoall function, and thus, require a large bisection bandwidth. Integer Sort (IS) , Fast Fourier Transformation (FT), Multigrid (MG), Conjugate Gradient (CG), Lower-Upper (LU) diagonal, Scalar Pentadiagonal (SP), Block Tridiagonal (BT), Embarrassingly Parallel (EP)

  19. EVALUATION USING PC CLUSTERS • NAS Parallel Benchmarks(NPBs) • “Tree (6link)” stands for the tree topology that uses six links between switches using link aggregation • “Compl(2link)” stands for a completely connected topology that uses two links between switches

  20. EVALUATION USING PC CLUSTERS • NAS Parallel Benchmarks(NPBs) • the proposed methodology improve the performance by up to 650 percent • Torus topology with three links between switches, that uses 36 links in total, achieves up to a 27 percent performance improvement compared with the tree topology with six links that uses 42 links in total.

  21. EVALUATION USING PC CLUSTERS • The simple static on/off link selection algorithm selects the deactivated links as follows: • 1) Estimate the amount of the application traffic in each channel when all the links are activated. The maximum amount of traffic on a single channel is calculated. • 2) Reduce the number of links between switches under the condition that the amount of the application traffic on every channel is less than the maximum amount of traffic on a channel calculated in step 1. the power consumption of all switches is reduced by up to 20 percent SuperNova cluster, 128 processes

  22. EVALUATION USING PC CLUSTERS • Impact of On/Off Link Regulation • the power consumption of all switches is reduced by up to 19 percent Misc cluster

  23. EVALUATION USING PC CLUSTERS • Impact of On/Off Link Regulation • the power consumption can be reduced by up to 25 percent in the case of PowerConnect 5324s Power consumption of on/off link regulation (SuperNova cluster, 64 processes, PC5324).

  24. CONCLUSIONS • Proposed a switch-tagged routing methodology • Implement the routing algorithms on PC clusters with Ethernet • Simple host configuration and high portability • Evaluation results • NAS parallel benchmarks performance comparable to that of an ideal 1-switch (full crossbar) network, • The torus topology achieves up to a 27 percent performance improvement compared with the tree topology using link aggregation. • The on/off and multispeed link regulation reduces the power consumption of switches by up to 25 percent

More Related